In a world of exponentially increasing information and countless forms of intelligent agents to process it, understanding the nature of intelligence - both natural and artificial - has become more necessary than ever. Intelligence is often conceived as monolithic: a universal organizing force that enables its host to solve hard problems and reshape its environment. And its various definitions often take an anthropomorphic character.
While we recognize that many living things have some sort of intelligence, humans, many of us believe, are placed at the top of the hierarchy. We hold that belief with a very vague definition of what that intelligence is, how it emerges, and where it is located - or if it even has a location. Although the definition doesn’t seem to matter at first, the moral code amongst the living often emerges based on implicit assumptions about the relationship between intelligence, consciousness, rights, and power. So asking the right questions about intelligence, especially as machines are taking over, is essential.
Let’s Start with The Living
How does a collection of living, breathing, cells self-organize into much bigger intelligent beings? Are cells intelligent themselves? And how do molecules self-organize into cells? Can they be called intelligent? And how do social systems and ecosystems emerge from elementary building blocks that are programmed to fend for themselves?
In this article, I argue that intelligence comes in many forms, characteristic of the diversity found in living and non-living organisms alike. Constraining it to be defined in terms of its most complex forms, as experienced by humans, misses the bigger picture of its purpose in nature, its ubiquity, and what it means for non-animal and algorithmic entities to be intelligent.
The best way to think about the nature of human and artificial intelligence is to first allow it to live in a higher dimensional space, by referring to it in the plural: intelligences. This is reminiscent of Gardner's theory of multiple intelligences; serving the same purpose to relax the reductionist assumption that intelligence is one-dimensional.
Within both natural and artificial intelligence, lives a zoo of intelligences, processes that are optimized to solve a variety of tasks. Those processes are coupled with each other to form greater intelligences. Some of them are rigidly and robustly designed, while others are constantly updated as a function of their feedback experiences.
The key aspect of intelligence is its multi-scale nature: most intelligences are formed of smaller intelligences and are part of bigger ones. Humans are made of cells, and well-functioning societies are made of intelligent humans. All systems that persist in time exist thanks to a spatially and temporally local stability that is adaptable and flexible enough to surrounding changes. From this perspective, intelligent systems optimize for stability through complex forms of coupling and decoupling with their surroundings, giving them access to larger scales of spatio-temporal stability.
Living things seem to go through cycles of stability of: birth, growth, death, and rest. The intelligence of living things heavily relies on the ability to balance the strategies of coupling and decoupling amongst other systems available to their input-output feedback loops to form more stable and powerful super-systems that maximize access to higher scales of spacial and temporal stability. More can be said about living things, but what about the intelligence of the artificial?
The Reductionist Fallacy of Intelligence
The recent obsession with general artificial intelligence (AGI) seems to miss the lesson we’ve learned from the successes of connectionist emergent learning over top-down rule-based intelligence. All evidence points to the fact that intelligence grows out of a collaborative self-organization of smaller intelligences, just like organisms grow out of smaller self-organized processes. While some intelligences are much more general than others, capable of performing a wider variety of tasks and self-organizing into even larger intelligences, they are all limited by the scale and scope of tasks they are optimized for. There is no such thing as a General Intelligence; unless, maybe, we are referring to the whole universe.
The pursuit of a general algorithm that can solve all super-human level tasks might sound novel, but it makes the same fallacy (or dream?) that science and religion have made for millennia: the reduction of complexity and diversity into a single entity that can answer all questions and solve all paradoxes. The desire is understandable, but with the limited means at our disposal, oversimplification is inevitable.
The Characteristics of Intelligences
There are many ways to characterize intelligence; and a frequently used metric is its ‘degree’. This is what the IQ test measures. While degree can be useful for evaluating a specific, well-defined, task, it is not very useful when considering intelligence as the ability to adapt to one’s environment. The tendency to overemphasize the degree exposes our belief that intelligence is context-independent and a host’s adaptation is only a function of a specific type of problem solv-ability. This is unsurprising since those who have the intelligence required to build and use tools have more power and control over their environments than others: they are more likely to survive and favor the type of intelligence that makes them superior.
An equally, if not more important measure of intelligence is its characteristic time-scale. Some intelligences change very quickly, updating with every single bit of information they are exposed to, while others are more static, changing every few years or even centuries. For example, the intelligence of metabolic processes that maintain the stability of a human body do not change that often over the centuries. Genes can be traced back millennia, while their collective organization into DNAs gets updated with every organism; staying constant throughout an organism’s lifetime. Brains are made to adapt quickly to a changing environment and learn quickly from limited experiences.
But even within brains, there are various intelligences that adapt to processes with different time-scales; some of them being learned over decades, while others being updated at every waking moment. These characteristic time-scales apply to algorithms as well: neural networks can adapt, learn and forget, while search algorithms are interpretable, robust and predictable.
Another important feature of intelligences is their ability to couple with each other. They do so through the hosts they control. For example, simple intelligences (e.g. algorithms) can be easier to couple with other intelligences than others. Intelligences that work well with others have the potential to self-organize into much more powerful intelligences. For example, one can think of neural networks as a collection of tiny intelligences (simple gates that synthesize multiple inputs into a single output) called neurons. They’re essentially simple input-output maps which can be easily composed into larger input-output maps that learn from input-output examples.
A similar logic can be applied to cells that self-organize into organs, and those into organisms. In other words, a powerful feature of an intelligence is to be a team-player so that it can build even more complex intelligences. This is, for example, the type of intelligence that a social organization (like a company) looks for in an employee. It is also the type of intelligence required for an efficient and functioning society.
Intelligences in Relationships
If we accept that intelligence is not restricted to those skills that manipulate symbols or solve complex tasks, our perspective on the relationship we have with other intelligent systems (animal or machine) might change.
For one, the relationship between intelligent systems that are required to solve a landscape of tasks is governed by a tension of collaboration and competition. When multiple intelligent agents have the same capabilities, they are prone to compete, and when they have complementary abilities, they are better off when they collaborate. If we create intelligent machines, both effects of competition and collaboration are to be expected. We can collaborate with computers that learn from data to create better science, but these systems can also compete with our jobs.
This tension also exists across scales: agents and the super-agents within which they exist might have conflicting goals. I might be interested in pursuing my passion in writing, but the community in which I exist will push me to contribute to its own growth in more pressing matters: like building machines and roads. Similarly, a soldier dies so that their country survives.
The complexity in the coupling of the various intelligences that exists within and across scales highlights the true dynamic and multifaceted nature of intelligence.
If intelligence is to be conceived as the ability of matter to self-organize into a stable entity that reshapes its environment, we must first get rid of the idea that it only lives in us, and that it only comes in limited forms; just as we dispelled the belief that life only exists at our scale.